Ev3Dev's picture
Upload folder using huggingface_hub
db03c40 verified
"""Decomposable reward function for the bio-experiment planning POMDP.
Reward components
─────────────────
r_validity β€” biological validity of the chosen action
r_ordering β€” correct ordering of experiment steps
r_info_gain β€” information gain from the step's output
r_efficiency β€” resource efficiency (budget & time normalised)
r_novelty β€” bonus for non-redundant, non-trivial actions
r_penalty β€” penalties for violations, redundancy, waste
r_terminal β€” terminal quality & calibration against hidden truth
Potential-based shaping
Ο†(s) β€” progress potential used for dense shaping signal
The final step reward is:
R_t = r_validity + r_ordering + r_info_gain + r_efficiency
+ r_novelty + r_penalty + [Ο†(s_{t+1}) βˆ’ Ο†(s_t)]
The terminal reward adds:
R_T += r_terminal
"""
from __future__ import annotations
from dataclasses import dataclass, field
from typing import Dict, List, Optional
from models import (
ActionType,
ConclusionClaim,
ExperimentAction,
IntermediateOutput,
META_ACTIONS,
TOOL_REGISTRY,
WET_LAB_ACTIONS,
)
from server.biology.gene_index import (
marker_set_score,
mechanism_set_score,
score_pathways,
)
from server.simulator.latent_state import FullLatentState
@dataclass
class RewardBreakdown:
validity: float = 0.0
ordering: float = 0.0
info_gain: float = 0.0
efficiency: float = 0.0
novelty: float = 0.0
penalty: float = 0.0
shaping: float = 0.0
terminal: float = 0.0
components: Dict[str, float] = field(default_factory=dict)
@property
def total(self) -> float:
return (
self.validity
+ self.ordering
+ self.info_gain
+ self.efficiency
+ self.novelty
+ self.penalty
+ self.shaping
+ self.terminal
)
def to_dict(self) -> Dict[str, float]:
d = {
"validity": self.validity,
"ordering": self.ordering,
"info_gain": self.info_gain,
"efficiency": self.efficiency,
"novelty": self.novelty,
"penalty": self.penalty,
"shaping": self.shaping,
"terminal": self.terminal,
"total": self.total,
}
d.update(self.components)
return d
class RewardComputer:
"""Computes step-wise and terminal rewards.
Parameters
----------
efficiency_weight : float
Relative importance of resource efficiency.
"""
def __init__(
self,
efficiency_weight: float = 0.3,
info_gain_weight: float = 0.4,
validity_weight: float = 0.3,
):
self.w_eff = efficiency_weight
self.w_ig = info_gain_weight
self.w_val = validity_weight
# ── step reward ─────────────────────────────────────────────────────
def step_reward(
self,
action: ExperimentAction,
prev_state: FullLatentState,
next_state: FullLatentState,
output: IntermediateOutput,
hard_violations: List[str],
soft_violations: List[str],
) -> RewardBreakdown:
rb = RewardBreakdown()
# validity
if hard_violations:
rb.validity = -1.0
rb.penalty = -0.5 * len(hard_violations)
rb.components["hard_violations"] = len(hard_violations)
return rb
rb.validity = self.w_val * (1.0 if output.success else 0.0)
ordering_score = self._ordering_score(action, prev_state)
rb.ordering = 0.2 * ordering_score
if ordering_score < 0:
rb.penalty += ordering_score * 0.3
# information gain proxy: quality Γ— (1 - uncertainty)
rb.info_gain = self.w_ig * output.quality_score * (1.0 - output.uncertainty)
if action.action_type in META_ACTIONS and not (
prev_state.progress.de_performed
or prev_state.progress.cells_clustered
):
# Meta actions before substantive analysis should not dominate reward.
rb.info_gain *= 0.2
# efficiency: normalised cost relative to budget
budget_frac = (
(next_state.resources.budget_used - prev_state.resources.budget_used)
/ max(next_state.resources.budget_total, 1)
)
rb.efficiency = self.w_eff * max(0.0, 1.0 - 5.0 * budget_frac)
# novelty: small bonus for non-redundant steps
if not soft_violations:
rb.novelty = 0.1
# tool-modality fit bonus/penalty
tool_fit = self._tool_fit_score(action, prev_state)
rb.components["tool_fit"] = tool_fit
rb.validity += 0.15 * tool_fit
# penalties
rb.penalty = -0.15 * len(soft_violations)
if action.action_type in META_ACTIONS and not (
prev_state.progress.de_performed
or prev_state.progress.cells_clustered
):
rb.penalty -= 0.25
rb.components["premature_meta_action_penalty"] = -0.25
# potential-based shaping (Ξ³=1 so it doesn't depend on the
# training algorithm's discount factor)
phi_prev = self._potential(prev_state)
phi_next = self._potential(next_state)
rb.shaping = phi_next - phi_prev
return rb
# ── terminal reward ─────────────────────────────────────────────────
def terminal_reward(
self,
state: FullLatentState,
conclusions: List[ConclusionClaim],
task_success_criteria: List[str],
discovered_markers: Optional[List[str]] = None,
candidate_mechanisms: Optional[List[str]] = None,
) -> RewardBreakdown:
rb = RewardBreakdown()
discovered_markers = discovered_markers or []
candidate_mechanisms = candidate_mechanisms or []
# pipeline completeness (0-1)
completeness = self._completeness(state)
rb.components["completeness"] = completeness
# calibration: how well conclusions align with hidden ground truth
calibration = self._calibration(state, conclusions)
rb.components["calibration"] = calibration
# efficiency bonus at terminal
budget_eff = state.resources.budget_remaining / max(
state.resources.budget_total, 1
)
time_eff = state.resources.time_remaining_days / max(
state.resources.time_limit_days, 1
)
rb.components["budget_efficiency"] = budget_eff
rb.components["time_efficiency"] = time_eff
# over-confidence penalty
overconf = self._overconfidence_penalty(state, conclusions)
rb.components["overconfidence_penalty"] = overconf
discovery_alignment = self._discovery_alignment(
state,
discovered_markers,
candidate_mechanisms,
)
discovery_error_penalty = -6.0 * (1.0 - discovery_alignment)
if discovery_alignment < 0.25:
discovery_error_penalty -= 2.0
rb.components["discovery_alignment"] = discovery_alignment
rb.components["discovery_error_penalty"] = discovery_error_penalty
conclusion_alignment = self._conclusion_alignment(state, conclusions)
conclusion_error_penalty = -4.0 * (1.0 - conclusion_alignment)
if conclusions and conclusion_alignment < 0.25:
conclusion_error_penalty -= 1.5
rb.components["conclusion_alignment"] = conclusion_alignment
rb.components["conclusion_error_penalty"] = conclusion_error_penalty
eff_bonus = (budget_eff + time_eff) / 2.0 if completeness >= 0.3 else 0.0
rb.terminal = (
3.0 * completeness
+ 4.0 * calibration
+ 1.0 * eff_bonus
+ overconf
+ discovery_error_penalty
+ conclusion_error_penalty
)
return rb
# ── helpers ─────────────────────────────────────────────────────────
def _ordering_score(
self, action: ExperimentAction, s: FullLatentState
) -> float:
"""Heuristic: 1.0 if natural next, 0.3 if acceptable, -1.0 if premature."""
at = action.action_type
p = s.progress
NATURAL_NEXT = {
ActionType.COLLECT_SAMPLE: not p.samples_collected,
ActionType.PREPARE_LIBRARY: p.samples_collected and not p.library_prepared,
ActionType.SEQUENCE_CELLS: p.library_prepared and not p.cells_sequenced,
ActionType.RUN_QC: p.cells_sequenced and not p.qc_performed,
ActionType.FILTER_DATA: p.qc_performed and not p.data_filtered,
ActionType.NORMALIZE_DATA: p.data_filtered and not p.data_normalized,
ActionType.CLUSTER_CELLS: p.data_normalized and not p.cells_clustered,
ActionType.DIFFERENTIAL_EXPRESSION: p.data_normalized and not p.de_performed,
ActionType.PATHWAY_ENRICHMENT: p.de_performed and not p.pathways_analyzed,
ActionType.MARKER_SELECTION: p.de_performed and not p.markers_discovered,
ActionType.VALIDATE_MARKER: p.markers_discovered and not p.markers_validated,
ActionType.SYNTHESIZE_CONCLUSION: (
p.de_performed or p.cells_clustered
) and not p.conclusion_reached,
}
if NATURAL_NEXT.get(at, False):
return 1.0
has_evidence = any([
p.cells_clustered, p.de_performed, p.trajectories_inferred,
p.pathways_analyzed, p.networks_inferred, p.markers_discovered,
])
if at in META_ACTIONS and not has_evidence:
return -1.0
return 0.3
def _potential(self, s: FullLatentState) -> float:
"""Progress potential Ο†(s) β€” counts completed milestones.
Returns 0.0 at terminal states so that the shaping signal
telescopes correctly over the episode.
"""
if s.progress.conclusion_reached:
return 0.0
p = s.progress
milestones = [
p.samples_collected,
p.library_prepared,
p.cells_sequenced,
p.qc_performed,
p.data_filtered,
p.data_normalized,
p.cells_clustered,
p.de_performed,
p.pathways_analyzed,
p.markers_discovered,
p.markers_validated,
p.conclusion_reached,
]
return sum(milestones) / len(milestones)
def _completeness(self, s: FullLatentState) -> float:
p = s.progress
core = [
p.samples_collected,
p.cells_sequenced,
p.qc_performed,
p.data_filtered,
p.data_normalized,
p.de_performed or p.cells_clustered,
p.conclusion_reached,
]
return sum(core) / len(core)
def _calibration(
self, s: FullLatentState, conclusions: List[ConclusionClaim]
) -> float:
"""Structured set-similarity calibration against hidden ground truth.
Uses pathway-weighted Gaussian similarity for markers, semantic
similarity for mechanisms, and activity-weighted matching for pathways.
Falls back to legacy substring matching when structured fields are empty.
"""
if not conclusions:
return 0.0
pred_markers = [g for c in conclusions for g in c.top_markers]
pred_mechs = [m for c in conclusions for m in c.causal_mechanisms]
pred_pathways = {
p: v for c in conclusions for p, v in c.predicted_pathways.items()
}
has_structured = bool(pred_markers or pred_mechs or pred_pathways)
if has_structured:
m_score = marker_set_score(pred_markers, s.biology.true_markers)
mech_score = mechanism_set_score(
pred_mechs, s.biology.causal_mechanisms
)
pw_score = score_pathways(pred_pathways, s.biology.true_pathways)
return 0.50 * m_score + 0.35 * mech_score + 0.15 * pw_score
return self._legacy_calibration(s, conclusions)
@staticmethod
def _legacy_calibration(
s: FullLatentState, conclusions: List[ConclusionClaim]
) -> float:
"""Substring-based calibration kept for backward compatibility."""
true_mechanisms = set(s.biology.causal_mechanisms)
true_markers = set(s.biology.true_markers)
score = 0.0
n = len(conclusions)
for c in conclusions:
claim_lower = c.claim.lower()
match = any(m.lower() in claim_lower for m in true_mechanisms)
marker_match = any(m.lower() in claim_lower for m in true_markers)
if match or marker_match:
score += 1.0
else:
score -= 0.3
return max(0.0, min(1.0, score / max(n, 1)))
_METHOD_TO_TOOL: Dict[str, str] = {
"scanpy.pp.calculate_qc_metrics": "Scanpy",
"scanpy.pp.filter_cells": "Scanpy",
"scanpy.pp.filter_genes": "Scanpy",
"scanpy.pp.normalize_total": "Scanpy",
"scanpy.pp.log1p": "Scanpy",
"scanpy.pp.highly_variable_genes": "Scanpy",
"scanpy.pp.neighbors": "Scanpy",
"scanpy.tl.leiden": "Leiden",
"scanpy.tl.louvain": "Louvain",
"scanpy.tl.rank_genes_groups": "Scanpy",
"scanpy.tl.paga": "PAGA",
"scanpy.tl.umap": "UMAP",
"gseapy.prerank": "Scanpy",
"gseapy.gsea": "Scanpy",
"10x_chromium": "CellRanger",
"NovaSeq": "CellRanger",
}
@staticmethod
def _tool_fit_score(
action: ExperimentAction, s: FullLatentState
) -> float:
"""Score how well the chosen tool matches the task modality.
Returns +1.0 for a perfect match, 0.0 if no tool specified,
-1.0 for a known tool used on an incompatible modality.
"""
method = action.method
if not method:
return 0.0
resolved = RewardComputer._METHOD_TO_TOOL.get(method, method)
tool_spec = TOOL_REGISTRY.get(resolved)
if tool_spec is None:
return -0.5
modality = getattr(s, "task_modality", None)
if not modality or not tool_spec.modalities:
return 0.0
if modality in tool_spec.modalities:
return 1.0
return -1.0
def _overconfidence_penalty(
self, s: FullLatentState, conclusions: List[ConclusionClaim]
) -> float:
"""Penalise high-confidence claims that disagree with ground truth.
Checks structured fields (top_markers, causal_mechanisms) first;
falls back to claim substring matching for backward compatibility.
"""
penalty = 0.0
true_markers_lower = {m.lower() for m in s.biology.true_markers}
true_mechs_lower = {m.lower() for m in s.biology.causal_mechanisms}
true_set = true_markers_lower | true_mechs_lower
for c in conclusions:
if c.confidence <= 0.8:
continue
has_structured = bool(c.top_markers or c.causal_mechanisms)
if has_structured:
marker_hit = any(
g.upper().strip() in {m.upper() for m in s.biology.true_markers}
for g in c.top_markers
)
mech_hit = any(
any(kw in m.lower() for kw in t.lower().split())
for m in c.causal_mechanisms
for t in s.biology.causal_mechanisms
)
is_correct = marker_hit or mech_hit
else:
is_correct = any(t in c.claim.lower() for t in true_set)
if not is_correct:
penalty -= 0.5 * c.confidence
return penalty
def _discovery_alignment(
self,
s: FullLatentState,
discovered_markers: List[str],
candidate_mechanisms: List[str],
) -> float:
"""Symmetric end-of-episode similarity for discovered biology.
Forward scoring measures recall against hidden truth. Reverse scoring
measures how well the agent's discoveries map back onto real biology,
which penalizes extra hallucinated markers or mechanisms.
"""
components: List[float] = []
if s.biology.true_markers or discovered_markers:
marker_recall = marker_set_score(
discovered_markers,
s.biology.true_markers,
)
marker_precision = marker_set_score(
s.biology.true_markers,
discovered_markers,
)
components.append((marker_recall + marker_precision) / 2.0)
if s.biology.causal_mechanisms or candidate_mechanisms:
mechanism_recall = mechanism_set_score(
candidate_mechanisms,
s.biology.causal_mechanisms,
)
mechanism_precision = mechanism_set_score(
s.biology.causal_mechanisms,
candidate_mechanisms,
)
components.append((mechanism_recall + mechanism_precision) / 2.0)
if not components:
return 1.0
return sum(components) / len(components)
def _conclusion_alignment(
self,
s: FullLatentState,
conclusions: List[ConclusionClaim],
) -> float:
if not conclusions:
return 0.0
pred_markers = [marker for conclusion in conclusions for marker in conclusion.top_markers]
pred_mechanisms = [
mechanism
for conclusion in conclusions
for mechanism in conclusion.causal_mechanisms
]
if not pred_markers and not pred_mechanisms:
return self._legacy_calibration(s, conclusions)
components: List[float] = []
if s.biology.true_markers or pred_markers:
marker_recall = marker_set_score(pred_markers, s.biology.true_markers)
marker_precision = marker_set_score(s.biology.true_markers, pred_markers)
components.append((marker_recall + marker_precision) / 2.0)
if s.biology.causal_mechanisms or pred_mechanisms:
mechanism_recall = mechanism_set_score(
pred_mechanisms,
s.biology.causal_mechanisms,
)
mechanism_precision = mechanism_set_score(
s.biology.causal_mechanisms,
pred_mechanisms,
)
components.append((mechanism_recall + mechanism_precision) / 2.0)
if not components:
return 1.0
return sum(components) / len(components)